DL_FFLUX Refactored: Accelerating Quantum Chemical Topology Simulations With Optimized Parallelization

Abstract

Molecular dynamics simulations are only as useful as they are both accurate and efficient. Here, we introduce an upgraded version of the simulation code DL_FFLUX that dramatically extends accessible time scales by combining a more efficient core algorithm with OpenMP and MPI parallelization. DL_FFLUX is a unique, state-of-the-art polarisable machine learning potential (MLP) that combines Quantum Chemical Topology (QCT) and machine learning (ML) and relies on DL_POLY as the driver for molecular dynamics simulations. This new implementation of DL_FFLUX delivers a massive performance boost, making simulations 6.8 to 49 times faster without compromising accuracy. This performance boost of DL_FFLUX is reported for condensed-phase simulations of water and ethanol and for single-molecule simulations of 6 small to medium-sized molecules. Moreover, we compared the optimized version of DL_FFLUX with other MLPs on the MD17 dataset for four molecules. DL_FFLUX is the fastest: depending on the system, it outperforms the second-fastest MLP by a factor of 1.8–9.6.